Abstract:Detecting surface defects on steel, especially in complex loading environments, poses significant challenges. In response, we introduce EDFW-YOLO, an algorithm built upon YOLOv8 specifically designed for detecting surface defects on hot-rolled steel strips. Our method enhances multi-scale feature fusion through the incorporation of the multi-scale con-version module C2f-EMSC. Additionally, we elevate detection accuracy by integrating the Dynamic Head target detection head, the Focal Modulation module, and the WIoU_Loss bounding box regression function. Experimental results on the NEU-DET dataset demonstrate that our optimized YOLOv8 model achieves an average accuracy (mAP) of 77.7%, with a 5.2% increase in network constraint rate. To adapt to different operating environments, it further improved the average accuracy (mAP) to 78.5% through data enhancement. Verification results on PCB defect data show that the algorithm has excellent generalization ability. This optimized algorithm significantly improves the extraction and fusion of surface defect features on hot-rolled strip steel and serves as a valuable reference for surface defect detection in alloy materials.